Retail & Hospitality /Document Generation live field guide · 8 min

Visual Product Search

Takes a photo of a product and searches the catalogue by image to return a ranked shortlist of matching SKUs with similarity scores and stock status — a staff member confirms the match before it drives an order or recommendation.

Open the live demo

theater/demos/retail-hospo_visual-search.html · sandbox · read-only

Open

FIG. 1

The live demo, running on fabricated data. Open it to step through the full flow — every output is shown for a person to approve before anything happens.

How it would work

Takes a photo of a product, extracts its visual features and matches them against the catalogue, then hands a ranked shortlist of SKUs — match score, price and stock — to a staff member to confirm before anything is quoted or ordered.

Input 01

A photo of the product

A customer's phone shot, a shop-floor photo, or an image pulled from an email — plus the retailer's product catalogue with images, SKUs, prices and current stock.

Agent 02

Extracts features, ranks SKUs

Reads category, colour, style and a brand-likeness score from the image, compares it to catalogue embeddings, and returns the closest SKUs ranked by visual similarity with a match percentage on each.

Output 03

A shortlist, for a person to confirm

A ranked list of candidate SKUs with match score, price, rating and stock, shown to a floor staffer or buyer who confirms the right item before it is quoted to a customer or sent to ordering.

Where it works well

It turns "I want the one in this photo" into a shortlist of real SKUs in seconds.

Best where customers routinely arrive with reference photos rather than codes — fashion, footwear, homewares, general merchandise.
Done by hand a hard-to-name item is minutes of scrolling per query; at volume that is hours a week of floor staff time.
The recaptured time goes back into the customer conversation — sizing, alternatives, the upsell — not the catalogue scroll.

The slow, invisible cost is the lookup — a customer arrives with a photo and no product code, and a staffer scrolls the catalogue guessing at colourways and ranges that look almost identical across a 500-plus SKU catalogue.

Where it works badly

It is confidently wrong when two SKUs look near-identical but differ in the detail that matters.

Weak on attributes a photo can't show — fabric, fit, material grade, the spec hidden inside a near-identical exterior.
A high match score is visual similarity, not a guarantee of the right item; treat it as a shortlist to confirm, never an answer to act on.
If your catalogue images are inconsistent — mixed lighting, missing angles, stock photos for half the range — match quality drops sharply.

The honest test

If a staffer would quote a price or place an order off the top result without opening the product page to confirm it, this tool makes that mistake faster, not safer.

Same silhouette, same colourway, different range, different price, or a genuine versus look-alike brand — the tool returns a clean 95% match that is the wrong SKU, and the score makes it look more certain than it is.

What it doesn't do — and shouldn't

It shortlists. A person confirms the SKU. That boundary is deliberate.

WHAT IT DOES

Surfaces the candidate SKUs ranked by visual similarity, with a match score on each

Shows the price, rating and current stock alongside each candidate

Names the visual features it matched on — category, colour, style, brand likeness

WHAT IT WON’T

Quote a price to the customer on its own

Place the order or commit stock

Decide that a look-alike is the genuine branded item

Quoting the wrong SKU or price to a customer is a false or misleading representation under the Australian Consumer Law (sections 18 and 29), which the ACCC enforces — and presenting a look-alike as a genuine branded product is its own exposure. The consequence lands on the retailer, so the accountable staffer stays on the confirm step; the tool only narrows the field.

What your data has to look like

A catalogue with clean, consistent product images and SKUs, priced and stocked in real time.

Typical readiness

across orgs we see, before the first job

Consistent catalogue images per SKU

Needs shaping

A structured SKU catalogue

Usual weak point

Live stock and current prices

Usual weak point

Product images for the long tail

Needs shaping

An image-search surface staff can reach

Usually ready

The real first job

The weak point is almost always the catalogue images — half the range shot inconsistently, the long tail not shot at all, prices and stock held in a system the images don't link to. Fixing how product images, SKUs, prices and stock are captured and kept current is usually the real first job — larger and more valuable than the matching layer on top. Once the catalogue is clean, every search after that is faster and right by default.

Right fit if…

Customers regularly arrive with photos rather than product codes or SKUs

A catalogue of 500-plus SKUs where ranges look visually similar

You already hold clean, consistent product images for most of the range

Floor staff or buyers lose real time to manual catalogue lookups

Walk away if…

Your catalogue images are inconsistent, incomplete, or mostly supplier stock photos

Products are reliably identified by barcode or code, so visual search adds nothing

The difference that matters is spec or fabric a photo can't show

You want it to quote prices or place orders without a person confirming

Open questions

The worried-buyer questions, answered straight

It can return a wrong match — which is exactly why nothing is quoted or ordered on its say-so. It shows a ranked shortlist with a match score, the price, the stock and the features it matched on, and a staff member confirms the SKU before anything reaches the customer. Quoting the wrong price is a false or misleading representation under the Australian Consumer Law, so the confirm step is the safeguard, not an optional extra.

Match quality follows image quality. The tool compares the customer’s photo to your catalogue images, so inconsistent lighting, missing angles or supplier stock photos all drag down the matches, and the long-tail SKUs you never photographed simply won’t appear. Getting clean, consistent images across the whole range is usually the first piece of work — and the piece that pays off across every search after.

No. It removes the catalogue scroll — the minutes of guessing at near- identical colourways — so the staffer spends that time on the customer: sizing, alternatives, whether the in-stock match actually suits them. The confirm and the conversation stay with the person; the freed capacity goes back into selling, not searching.

Current enough that the shortlist reflects what you can actually sell today. Prices and stock change daily in retail, and a match that shows an item as available at last week’s price sets up a wrong quote to the customer. The images need to cover the live range; the price and stock behind each SKU need to be today’s, not a stale export.

A customer photo and your catalogue, pricing and stock are commercial and, where a person is identifiable in a shot, personal information under the Privacy Act. Any deployment runs against your own systems and data handling, not a shared pool — we scope where the images sit and who can see them as part of the build. The demo here runs entirely on fabricated data; the products and stock shown are not real.

No — it reports a brand-likeness score from the visual features, not an authentication. A look-alike can score high, and presenting it as the genuine branded product is the retailer’s exposure, not the tool’s. The score narrows the shortlist; a person confirms the actual SKU and brand before quoting or ordering.

What it takes to build

4–6 weeks · 4 phases

Reused from template~65%

Bespoke to this skin~35%

stack · Claude vision · image embeddings · inventory connectors · review UI

What it would cost

Fixed scope, fixed price, fixed dates.

Bite-sized first piece

One contained change, low risk

Pilot build

Most builds land here

Embedded support

Scale on proof

Considering this for your store?

The honest place to start is a bite-sized first piece — one contained change, low risk. Tell us where it hurts; we'll play it back, scope it, and show you what's possible.

Book a call How we work